File size: 7,181 Bytes
8ddfad6
 
 
 
 
 
 
 
 
 
 
 
 
49234f8
8ddfad6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49234f8
 
 
 
 
 
 
8ddfad6
 
 
 
49234f8
8ddfad6
49234f8
8ddfad6
 
 
 
 
 
 
 
 
 
 
a2d30b7
 
8ddfad6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a2d30b7
 
8ddfad6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a2d30b7
 
8ddfad6
 
49234f8
8ddfad6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49234f8
 
8ddfad6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49234f8
 
8ddfad6
49234f8
 
 
 
 
 
 
 
 
 
 
 
8ddfad6
 
 
 
 
 
 
 
 
 
 
 
 
 
49234f8
 
8ddfad6
 
 
49234f8
 
8ddfad6
 
 
 
49234f8
 
8ddfad6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
name: document-classification-env
version: 1.0.0
description: >
  A real-world document classification and routing environment where AI agents 
  learn to classify documents and route them to appropriate departments with 
  optimal speed and accuracy tradeoffs.

author: OpenEnv Community
license: MIT

# Environment specification
environment:
  id: DocumentClassification-v1
  entry_point: server.document_classification_environment:DocumentClassificationEnvironment
  
  # Observation space definition
  observation_space:
    type: dict
    properties:
      document_id:
        type: string
        description: Unique identifier for the document
      content:
        type: string
        description: Raw text content of the document
      word_count:
        type: integer
        description: Number of words in the document
        min: 0
      has_urgency_markers:
        type: boolean
        description: Whether document contains urgency indicators
      features:
        type: array
        items:
          type: number
        description: Pre-extracted TF-IDF features (normalized)
        shape: [100]  # 100-dimensional feature vector
      metadata_response:
        type: string
        description: Output from the Request Metadata tool
      sla_remaining:
        type: number
        description: Seconds remaining before SLA breach for current document
        min: 0
  
  # Action space definition
  action_space:
    type: discrete
    description: Category selection or Tool Usage. Actions 0 to N-1 classify the document. Action N runs 'Request Metadata' tool. Action N+1 runs 'Escalate to Human'.
    # Values depend on difficulty level
    # Easy: 7 actions, Medium: 12 actions, Hard: 24 actions

# Tasks definition
tasks:
  - id: easy
    name: Document Classification - Easy
    difficulty: easy
    description: >
      Classify pre-processed documents into 5 categories with no time constraints.
      Documents are pre-extracted and normalized. This task serves as an introduction
      to the environment and API.
    
    grader: ".grading:AgentGrader"
    
    parameters:
      num_documents: 100
      num_categories: 5
      categories:
        - General
        - Billing
        - Support
        - Technical
        - HR
      time_limit_per_action: null
      time_limit_per_episode: null
    
    grading:
      type: accuracy_based
      success_threshold: 0.95
      test_dataset_size: 100
      metrics:
        - accuracy
        - precision_per_category
        - recall_per_category
  
  - id: medium
    name: Document Classification - Medium
    difficulty: medium
    description: >
      Classify raw text documents into 10 categories with moderate complexity.
      Includes time constraints per action. Requires the agent to balance
      processing time with accuracy.
    
    grader: ".grading:AgentGrader"
    
    parameters:
      num_documents: 500
      num_categories: 10
      categories:
        - General
        - Billing
        - Billing-Dispute
        - Support
        - Technical
        - Technical-Bug
        - HR-Payroll
        - HR-Benefits
        - Legal
        - Executive
      time_limit_per_action: 2.0  # seconds
      time_limit_per_episode: 1800  # 30 minutes for full dataset
    
    grading:
      type: accuracy_with_time_bonus
      success_threshold: 0.85
      test_dataset_size: 500
      metrics:
        - accuracy
        - average_response_time
        - speed_efficiency_ratio
        - precision_per_category
        - recall_per_category
  
  - id: hard
    name: Document Classification - Hard
    difficulty: hard
    description: >
      Classify complex documents into 20 fine-grained categories with 
      challenging real-world scenarios. Tight time constraints create
      speed-accuracy tradeoffs. Documents may contain mixed formats
      and ambiguous classification cases.
    
    grader: ".grading:AgentGrader"
    
    parameters:
      num_documents: 1000
      num_categories: 22
      categories:
        - General
        - Billing
        - Billing-Dispute
        - Billing-Refund
        - Support-Urgent
        - Support-Normal
        - Technical
        - Technical-Bug
        - Technical-Feature
        - HR-Payroll
        - HR-Benefits
        - HR-Complaint
        - Legal
        - Legal-Contract
        - Legal-Compliance
        - Executive
        - Executive-Strategic
        - Finance
        - Marketing
        - Operations
        - Finance-Tax
        - Marketing-Social
      time_limit_per_action: 1.0  # seconds
      time_limit_per_episode: 1200  # 20 minutes for full dataset
    
    grading:
      type: accuracy_with_speed_cost
      success_threshold: 0.75
      test_dataset_size: 1000
      metrics:
        - accuracy
        - response_time
        - speed_accuracy_tradeoff_score
        - confusion_matrix
        - f1_score_macro
        - f1_score_weighted

# Reward function specification
reward:
  type: composite
  description: >
    Composite reward combining accuracy, speed, and penalizing errors.
    The exact calculation depends on task difficulty.
  
  components:
    - name: accuracy_bonus
      description: Positive reward for correct classification
      weight: 0.7
      value_range: [0, 1]
    
    - name: speed_bonus
      description: Bonus reward for fast decision making
      weight: 0.2
      value_range: [0, 0.2]
      scales_with: response_time_efficiency
    
    - name: error_penalty
      description: Penalty for incorrect classification
      weight: 0.1
      value_range: [-0.5, 0]
    
    - name: processing_cost
      description: Small deduction (-0.01) for resource usage per action (hard task only)
      weight: 0.1
      task_specific: hard
      value_range: [-0.01, 0]
      
    - name: tool_usage_cost
      description: Deduct -0.05 reward when agent requests metadata to penalize unnecessary tool usage.
      weight: 0.05
      value_range: [-0.05, 0]

    - name: sla_penalty
      description: Time-pressure penalty that scales with how long the agent deliberates on a document beyond the SLA window. Caps at -0.3.
      weight: 0.1
      value_range: [-0.3, 0]
      task_specific: medium, hard

# Episode structure
episode:
  max_steps: null  # No step limit, episode ends when all documents processed
  reset_regenerates_data: true
  observation_includes_progress: true
  done_conditions:
    - all_documents_classified
    - time_limit_exceeded
    - manual_termination

# Baseline performance targets
baseline_targets:
  easy:
    score: 0.80
    accuracy: 0.80
    time_per_action_ms: 50
  
  medium:
    score: 0.79
    accuracy: 0.79
    time_per_action_ms: 150
    speed_bonus: 0.15
  
  hard:
    score: 0.23
    accuracy: 0.23
    time_per_action_ms: 100
    accuracy_std_dev: 0.05

# Reproducibility settings
reproducibility:
  seed_controllable: true
  deterministic_data_generation: true
  deterministic_grading: true
  fixed_test_sets: true

# Deployment settings
deployment:
  containerized: true
  docker_image: doc-classifier-env:latest
  huggingface_spaces: true
  api_endpoint: /api/step
  inference_framework: gymnasium_compatible